面向视频结构化的细粒度车辆检测分类模型

In order to solve the problem of limited understanding of complex traffic scenes in driverless environment perception technology, this paper proposes a roadside-oriented video structured description framework, which can enrich the fine-grained information of different targets in traffic scenes and improve the understanding ability of complex traffic scenes. For the proposed framework, this paper provides an engineering fine-grained vehicle detection and classification model. The YOLOv4 algorithm is optimized by channel pruning strategy, and the volume of the compressed model, YOLOv4-Pruned, is reduced by about 60% compared with the original model under the condition that mAP is almost unchanged. A vehicle classification method with 16 types and 12 colors is designed, which can effectively cover all vehicles in the current traffic scene. And the classification accuracy of the test set can reach 93%. The fine-grained vehicle detection and classification model designed in this paper is stable at 23FPS under 1920 × 1080 pixel input, NVIDIA Geforce RTX 2080ti, and the unquantified model is stable at 13FPS under Hisilicon-Hi3516DV300.